grackle-in-a-HEB-parking-lot/possumm-local

POSSUMM Neumes: Musical Symbols from Medieval Music Manuscripts

⚠️ DATASET UNDER CONSTRUCTION - This is currently a private working repository serving as a data bucket while transitioning from GitHub storage. Structure and organization are actively being refined.

Dataset Description

The POSSUMM (Paleographic Object Script Sorter Using Machine Methods) Neumes dataset is a specialized collection of medieval musical symbols extracted from historical manuscripts. This dataset focuses on the automatic classification of medieval musical script types, supporting four distinct notation systems used across different regions and time periods.

🚧 UNDER CONSTRUCTION: Currently reorganizing from unstructured data dump - final organization pending!

Dataset Summary

This dataset contains neumes (medieval musical notation symbols) extracted from digitized manuscripts dating from the 10th to 15th centuries. Each neume is annotated with its script type, source manuscript information, and spatial coordinates. The dataset serves as a foundation for training machine learning models to automatically classify and recognize medieval musical notation.

Unlike the possumm-global dataset (also part of this project), this local dataset features specific neumes extracted from pages with relevant annotations connecting them to their type, page, and manuscript source.

Supported Script Types

The premise of the project is the automatic classification of medieval musical script types. Currently four script types are supported:

• San Gall (Class 0): Notation from the St. Gallen monastery tradition
• Hispanic (Class 1): Visigothic-Mozarabic notation from the Iberian Peninsula
• Beneventan (Class 2): Southern Italian notation style
• Square (Class 3): Later standardized notation

Extraction Methods

The majority of these neumes were extracted following one of the following methods:

• Hand annotation in the Visual Geometry Group's Image Annotator (VIA)
• Extracted using bounding box coordinates from MEI files produced as part of the end-to-end optical music recognition process conducted by the DDMAL Lab
• Converted and extracted from neume-specific annotation and description as part of the Optical Neume Recognition Project (Hankinson et al.)

Dataset Statistics

Based on the POSSUMM paper, the dataset comprises:

Global Dataset (Full Manuscript Pages)

• Total Images: 2,492 medieval manuscript images
• St. Gall: 986 images (Class 0)
• Hispanic: 432 images (Class 1)
• Beneventan: 385 images (Class 2)
• Square: 394 images (Class 3)
• Ambiguous/Other: 295 images (Class 4)

Local Dataset (Individual Neumes)

• Total Neumes: 39,966 extracted neume symbols
• St. Gall: 8,172 neumes
• Hispanic: 10,000 neumes
• Beneventan: 9,000 neumes
• Square: 12,794 neumes

Technical Specifications

• Image Format: Resized to 224×224 pixels, normalized for contrast
• Neume Types: Focus on paleographically distinctive forms (clivis, torculus, porrectus, liquescents, custos)
• Augmentation: +4,000 additional samples through historically-informed augmentation

Dataset Structure

🚧 UNDER CONSTRUCTION: Current structure is unorganized - need to document:

• Directory structure
• File naming conventions
• Annotation formats

Data Fields

• Neume images (224×224 pixels, normalized)
• Script type classifications (0-4)
• Manuscript source information
• Spatial coordinates/bounding boxes
• Extraction method metadata
• Neume type labels (clivis, torculus, porrectus, etc.)

Data Splits

🚧 UNDER CONSTRUCTION: The paper mentions 5-fold cross-validation - need to document final train/validation/test splits for public release.

Performance Benchmarks

Based on the POSSUMM dual-pathway neural network:

Classification Accuracy

• Overall Accuracy: 96% on test set
• St. Gall: F1 = 0.99 (Precision: 1.00, Recall: 1.00)
• Hispanic: F1 = 0.995 (Precision: 1.00, Recall: 0.99)
• Beneventan: F1 = 0.927 (Precision: 0.88, Recall: 0.98)
• Square: F1 = 0.964 (Precision: 0.98, Recall: 0.95)
• Ambiguous: F1 = 0.815 (Precision: 0.88, Recall: 0.76)

Comparison with Baseline Approaches

Model	Accuracy	Macro F1-Score
Single-stream VGG16	0.89	0.87
ResNet50	0.91	0.90
Dual-pathway (POSSUMM)	0.96	0.94

Source Manuscripts

All manuscripts used in this study are open source and accessible, with relevant citations and URLs below:

San Gall Sources (Class 0)

Source: https://e-codices.unifr.ch/en/list/subproject/stgall_music

Cod. Sang. 361

• Location: St. Gallen, Stiftsbibliothek, Switzerland
• Title: Gradual
• Date: First half of 12th century
• Description: Parchment · 142 pp. · 29.5-30 x 19.5 cm
• Language: Latin
• Summary: Gradual from St. Gall with finely executed neumes and illuminated initials. Includes Calendar with necrological notes (13th-15th century) and catalogue of relics (14th century).
• DOI: 10.5076/e-codices-csg-0361
• IIIF Manifest: https://e-codices.unifr.ch/metadata/iiif/csg-0361/manifest.json
• Citation: St. Gallen, Stiftsbibliothek, Cod. Sang. 361: Gradual (https://e-codices.unifr.ch/en/list/one/csg/0361)

Cod. Sang. 374

• Location: St. Gallen, Stiftsbibliothek, Switzerland
• Title: Gradual and lectionary with epistles and Gospels
• Date: Middle of 11th century
• Description: Parchment · 845 pp. · 21 x 16 cm
• Summary: Deluxe manuscript for feast day masses with gradual containing neumes and Lectionary with liturgical year readings
• DOI: 10.5076/e-codices-csg-0374
• IIIF Manifest: https://e-codices.unifr.ch/metadata/iiif/csg-0374/manifest.json
• Citation: St. Gallen, Stiftsbibliothek, Cod. Sang. 374: Gradual and lectionary with epistles and Gospels (https://e-codices.unifr.ch/en/list/one/csg/0374)

Cod. Sang. 388

• Location: St. Gallen, Stiftsbibliothek, Switzerland
• Title: Antiphonary
• Date: 12th century (additions until late 14th century)
• Description: Parchment · 498 pp. · 25-25.5 x 15.5-16 cm
• Summary: Antiphonary for divine office liturgy with miniature of crucified Christ with Mary and John
• DOI: 10.5076/e-codices-csg-0388
• IIIF Manifest: https://e-codices.unifr.ch/metadata/iiif/csg-0388/manifest.json
• Citation: St. Gallen, Stiftsbibliothek, Cod. Sang. 388: Antiphonary (https://e-codices.unifr.ch/en/list/one/csg/0388)

Hispanic Sources (Class 1)

Leon Antiphoner Ms 8

• Title: Liber antiphonarium de toto anni circulo a festivitate sancti Aciscli usque in finem
• Date: 10th century
• Description: 306 h. : perg., il. ; 33 x 24 cm
• Type: Visigothic-Mozarabic Antiphonary
• URL: https://bvpb.mcu.es/es/consulta/registro.cmd?id=449895
• References: Susana Zapke. Hispania Vetus: manuscritos litúrgico-musicales (2007); Agustin Millares Carlo. Manuscritos visigóticos (1963)

P-Cua Coimbra IV-3ª Gav. 44 (22)

• URL: https://pemdatabase.eu/source/46581
• Status: 2 images, deprecated/damaged

Liber Misticus

• Date: 10th century
• Description: 230 h. : perg. ; 38 x 28 cm
• URL: https://bvpb.mcu.es/es/consulta/registro.cmd?id=409852

Beneventan Sources (Class 2)

Roma Vallicelliana C32

• Location: Biblioteca Vallicelliana, Rome
• Date: 1051-1100
• Contributor: Stazio, Achille (possessore)
• URL: https://www.internetculturale.it/jmms/iccuviewer/iccu.jsp?id=oai%3Awww.internetculturale.sbn.it%2FTeca%3A20%3ANT0000%3ACNMD%5C%5C0000016295

Bodleian Library MS Canon. Liturg. 342

• Location: Oxford, Bodleian Library
• Date: 13th century, second half
• Origin: Dalmatia, Ragusa (Dubrovnik), Benedictine abbey of St Mary
• Description: 250 × 168 mm, Beneventan script and notation
• Special Notes: Contains mutilated pages, smeared ink, damaged sections - excellent for training reject models
• URL: https://digital.bodleian.ox.ac.uk/objects/11d1899d-ab98-4d5c-a940-ef8cd9d7d710/

P-Cug Coimbra MM 1063 (63)

• Type: Missal fragment
• Date: 13th century, 2nd half
• Notation: Late-style Beneventan with F- and C-clefs, red four-line staves
• URL: https://pemdatabase.eu/source/97169

Square Notation Sources (Class 3)

Christ Church MS 87

• Location: Oxford, Christ Church
• Date: 1400-1450
• Origin: England, London
• Type: Missal (Use of Sarum)
• Musical Script: Black square notation on red four-lined staves
• URL: https://digital.bodleian.ox.ac.uk/objects/d6070820-9e6e-4443-bfd6-8c269b3306f0/
• IIIF: https://iiif.bodleian.ox.ac.uk/iiif/manifest/d6070820-9e6e-4443-bfd6-8c269b3306f0.json

Plimpton Antiphonal MS 041

• Location: Columbia University, Rare Book & Manuscript Library
• Date: 1473
• Origin: Perugia, Italy
• Description: 171 leaves on parchment
• Contributors: Caporali, Giacomo (Illustrator), Alvise, Don (Scribe)
• URL: https://dlc.library.columbia.edu/catalog/ldpd:113554

McGill MS Medieval 0073

• Location: Montreal, McGill University Rare Books and Special Collections
• Date: 1275-1315 & 1400-1500
• Origin: Northern Italy
• Type: Antiphonary (blend of office and mass materials)
• Musical Script: Square notation, 6 identified musical scribes
• URLs: https://cantus.simssa.ca/manuscript/680970 | https://archive.org/details/McGillLibrary-rbsc_ms-medieval-073-18802

Einsiedeln Codex 611

• Location: Einsiedeln, Stiftsbibliothek, Switzerland
• Title: Antiphonarium pro Ecclesia Einsidlensi
• Date: Early 14th century (1300-1313)
• Description: Parchment, 280 folios, 32x22cm
• Musical Script: Black square notation on four-lined staves (usually red)
• URL: https://www.e-codices.ch/en/list/one/sbe/0611

Usage Examples

Data Loading

🚧 UNDER CONSTRUCTION: Add code examples showing:

# Placeholder - need to add actual loading code
from datasets import load_dataset

# Load global dataset (full manuscript pages)
global_dataset = load_dataset("possumm/neumes", "global")

# Load local dataset (individual neumes)  
local_dataset = load_dataset("possumm/neumes", "local")

Model Training

The paper demonstrates a dual-pathway Siamese network architecture:

• Global Branch: ResNet50 backbone for document-level features
• Local Branch: ResNet50 backbone for neume-level features
• Fusion Module: Concatenation + attention mechanism
• Training Strategy: Progressive unfreezing (3 phases)
• Optimization: Adam optimizer, categorical cross-entropy loss
• Hardware: Successfully trained on consumer hardware (Mac Mini M4, MacBook Pro M1)

Technical Implementation

Data Augmentation Strategy

Historically-informed transformations reflecting actual scribal variation:

• Rotational: Reflecting manuscript positioning variations
• Obscurational: Simulating damage and wear
• Noise: Reflecting parchment texture and ink flow variation
• Intensity Scaling: More aggressive for minority classes

Class Balancing

• Class Weights: Inversely proportional to sample frequency
  • St. Gall: 0.51
  • Hispanic: 1.15
  • Beneventan: 1.29
  • Square: 1.26
  • Ambiguous: 1.69

Interactive Disambiguation Protocol

For cases with confidence < 0.65:

• Bayesian Query-by-Example: Iterative user feedback
• Average Interactions: 3.2 to reach high confidence
• Success Rate: 83% of ambiguous cases resolved
• Information Gain: Adaptive sampling targeting diagnostic features

Applications and Use Cases

Digital Library Enhancement

• Metadata Enrichment: Automatic script classification for under-cataloged collections
• Cross-Collection Discovery: Enable queries by notation type across repositories
• Integration Points: IIIF workflows, batch processing, interactive verification

Pedagogical Applications

• Educational Tool: Interactive disambiguation teaches diagnostic features
• Research Support: Confidence scores and feature importance for paleographic analysis
• Accessibility: Democratizes specialized paleographic knowledge

Cultural Heritage Integration

• Linked Data: Compatible with IIIF, MEI, Wikidata frameworks
• Authority Integration: Connect with existing manuscript databases
• Semantic Enrichment: Machine-actionable metadata for advanced querying

Known Issues & Limitations

Based on the paper's findings:

Current Limitations

• Class Imbalance: St. Gall samples significantly outnumber other classes
• Regional Coverage: Limited to Western European neume traditions
• Boundary Cases: 17% of ambiguous cases remain challenging even for experts
• Fragment Quality: Performance may vary with manuscript condition

Specific Challenges

• Document-Level Similarities: Early notations (St. Gall/Beneventan) can overlap visually
• Scribal Variation: High variability within traditions due to individual practices
• Damage Handling: While robust to simulated damage, real degradation varies

Future Extensions Planned

• Additional Scripts: Aquitanian, Breton, Norman-Anglo notation
• Sub-Classifications: Roman-Beneventan, Hufnagl-Square variants
• Codebook Development: Tradition-specific recognition modules

Ethical Considerations

All manuscripts used are open-access and properly attributed to their holding institutions. The dataset respects cultural heritage preservation practices and supports scholarly research in digital paleography. No personal or sensitive data were included, and metadata has been cited according to source repository guidelines.

Citation

@dataset{bouressa2024possumm,
  title={POSSUMM Neumes: Musical Symbols from Medieval Music Manuscripts},
  author={Bouressa, Kyrie},
  institution={McGill University, Distributed Digital Media Archives and Libraries Lab (DDMAL)},
  year={2024},
  note={Dataset under construction - private repository}
}

@inproceedings{bouressa2025possumm,
  title={From Pixels to Paleography: A Dual-Pathway Neural Network for Neume Script Classification},
  author={Bouressa, Kyrie},
  booktitle={Proceedings of DFLM'25},
  year={2025},
  note={Unpublished working draft}
}

Development Notes

Current Status: Transitioning from GitHub storage to proper Hugging Face organization. Data is currently unstructured and being reorganized for public release.

Technical Achievement: 96% classification accuracy with dual-pathway architecture, successfully trained on consumer hardware, demonstrating accessibility for researchers without specialized computing resources.

Next Steps:

• Complete file structure organization
• Finalize train/validation/test splits
• Add comprehensive usage examples
• Integrate interactive disambiguation interface
• Prepare for public release

Contact

Kyrie Bouressa
McGill University
Distributed Digital Media Archives and Libraries Lab (DDMAL)

---

This dataset is part of ongoing research in digital paleography and optical music recognition for medieval manuscripts. All "🚧 UNDER CONSTRUCTION" sections will be completed during the reorganization phase.